The CSDC database covers 6 provinces, 41 hospitals and 12 population cohorts in China (Yu et al., 2016). The CSDC database facilitates stroke-related decision-making, research, and public health services through a comprehensive system. It collects and analyzes patient data, including risk factors, medical history, and sociodemographic information, ensuring that each subject has a unique record. A two-stage stratified cluster sampling method was employed during the data screening process (Li et al., 2019a). First, more than 200 screening areas were selected based on the local population size and the total number of counties. Then, urban communities and townships were used as the primary sampling units (PSUs) according to the geographical location and the recommendations from the local hospitals. In each PSU, all residents aged 40 and above were surveyed using cluster sampling during the initial screening period. Doctors assessed each patient’s condition, categorizing them as low risk, medium risk, high risk, transient ischemic attack (TIA), or stroke. The CSDC dataset comprises 862,244 middle-aged residents. Table 1 shows the detailed features of the CSDC dataset.

The in-house data is sourced from the medical records of 49 patients at affiliated hospital of Jining Medical University in 2023. It includes 14 features such as gender, age group, ethnic groups, marital status, occupation, education level, hypertension, atrial fibrillation, smoking, hyperlipidemia, diabetes, overweight, and family history of stroke. Each patient has been diagnosed by a physician and classified as either having suffered a stroke or being in good health. The summary information for these two datasets is listed in Table 2.

The SRPNet model mainly consists of two modules: two-level feature selection, and deep fusion network. The overall framework is illustrated in Figure 1. Since the dataset contains text information, the SRPNet firstly performs data preprocessing, which involves digitizing the textual information and normalizing the data. To eliminate low-correlation and redundant features, the two-level feature selection method is employed to identify comprehensive features associated with stroke. Finally, the deep fusion network, which adaptively fuse Transformer and FCN by attention mechanism, takes the obtained significant features as input to provide accurate stroke risk prediction results for stroke patients.

Based on the stroke risk researches (Tian et al., 2019; Guan et al., 2019), we used text information digitization to convert non-numeric features into numeric vectors suitable for machine learning or deep learning methods. Occupations are divided into mental workers and manual workers. For the marital status, we characterize it by whether the respondent is currently married and the number of marriage times. Based on the location information of the respondents, we convert it to the local climate, such as maximum temperature, minimum temperature, precipitation, humidity, etc. All of which are closely related to stroke. For the remaining features, we also use similar knowledge-based feature engineering for feature representation. Data normalization (Park et al., 2022) is used to scale data elements to the (0,1) interval, which helps improve the effectiveness and reliability of model training. The normalization formula is defined as follows Equation 1: x ′ = x − min x max x − min x . (1)

In this section, the two-level feature selection method that contains two-step feature selection processes will be introduced. The first step of feature selection involves four distinct methods, which are Pearson correlation, chi-square test, Lasso, and elastic net. The union of selected features from each method forms a set of candidate features. In the second step, based on the seven machine learning models, such as SVM, KNN, GBDT, RF, DT, AdaBoost and GaussianNB, we evaluate all possible combinations of candidate features via grid search. Each combination is scored based on its performance in the given models. It allows us to determine the optimal combination of features that are most predictive of stroke risk.

The two-step approach provides a rigorous feature selection process by multiple machine learning methods. The first step reduces the number of features based on statistical tests of relevance. The second step further refines the features by evaluating prediction performance in representative machine learning models. This ensures that the most informative and generalizable features have been selected for predicting stroke risk.

The complexity and diversity of stroke data require predicting stroke risk from multiple perspectives to enhance model robustness. Common predictive models, such as the Transformer, exhibit complex structures and excel at adapting to high-dimensional data, thus improving prediction performance. However, it often suffers from overfitting issues when dealing with small-scale datasets. In contrast, the FCN model has a simple structure and fast training speed, yielding exceptional performance on small-scale datasets. We utilize the advantages of both above predictive models and propose a deep fusion network method that can provide accurate the stroke risk prediction. As shown in Figure 2, deep fusion network integrates the Transform and the FCN, in which the dependencies between stroke risk factors are captured by the attention mechanism of the Transformer, and the complex nonlinear relationship is fitted by deep network structure of the FCN.

Transformer (Vaswani et al., 2017). Due to the powerful representation ability, Transformer can realize the outstanding performance in prediction tasks which is based on the self-attention mechanism. As observed in Figure 2, given an input X ∈ ℝ n × c , where n represents the number of patients (or patches) and c represents the embedded feature dimension for every patient. The self-attention mechanism can be defined as Equation 6: A t t e n t i o n Q , K , V = softmax Q K T d k V , (6) where d k is the input matrix embedding dimension. The matrix Q , K , V can be computed by the input matrix and linear transformation matrix W Q , W K , W V , respectively. Then, we can get the value of Q , K , and V by computing Q = X W Q , K = X W K , and V = X W V , where W Q ∈ ℝ c × q , W K ∈ ℝ c × q , W V ∈ ℝ c × q , q denotes the linear mapping dimension.

Fully connected neural network. The FCN, also known as a Multilayer Perceptron (MLP), is a widely used artificial neural network structure in medical data analysis. It offers the advantages of fast training speed and robust modeling capabilities as the network depth increases. The stroke risk prediction model we designed includes one input layer, one hidden layer and one output layer. The calculation formula of each layer of network is defined by Equation 7: y = σ W x + b , (7) where σ denotes the ReLU activation function. x is the input of the neuronal node, y is the output of the neuronal node. W and b denote weight and bias, respectively, which are learnable parameters.

Attention mechanism. In the stroke risk prediction task, the Transformer and the FCN extract clinical features at different levels and make distinct contributions to the prediction. Therefore, we introduce an attention mechanism to adaptively learn the importance of latent embeddings. Specifically, for the feature H t extracted by Transformer, we apply a non-linear transformation and employ the shared attention vector W t to obtain the attention coefficient a t , namely, Equation 8: a t = softmax W t ⋅ σ W H t + b , (8) where σ denotes the tanh activation, W denotes a trainable weight matrix, and b denotes a bias vector. Similarly, we can calculate attention coefficients a c for the features H c extracted by the FCN. We combine these embeddings to obtain the final embedding H , Equation 9: H = L a t ⋅ H t + a c ⋅ H c , (9) where L denotes the single linear layer.

Here we employ four evaluation metrics to assess the predictive performance of the model, including micro precision, micro F1-score, macro precision, and Cohen’s Kappa coefficient (Younas et al., 2023). The definitions of these metrics are given as follows.

The micro average approach amalgamates performance measures across all samples. Specifically, for each class g i within the set G = 1 , . . . , K , where K denotes the total number of classes, a dedicated confusion matrix is constructed. In this context, the i -th matrix designates the g i class as the positive class, while considering the remaining classes g j with j ≠ i as the negative classes. The micro precision and micro F1-score are computed by Equations 10 and 11: P m i c r o = ∑ i = 1 G T P i ∑ i = 1 G T P i + F P i , (10) F 1 m i c r o = 2 ∑ i = 1 G T P i 2 ∑ i = 1 G T P i + ∑ i = 1 G F P i + ∑ i = 1 G F N i , (11) where TP represents the number of positive samples correctly predicted to be positive samples, FP represents the number of negative samples incorrectly predicted to be positive samples, FN represents the number of positive samples incorrectly predicted to be negative samples.

Micro average tends to provide misleading results in the case of imbalanced data, as it doesn’t take the predictive performance of each specific class into account. In contrast, macro average computes averages through the individual performance of each class. The macro precision is defined as Equation 12: P m a c r o = 1 G ∑ i = 1 G T P i T P i + F P i . (12)

Cohen’s Kappa Coefficient is employed for assessing performance in situations of imbalanced class distribution, which is denoted by Equation 13: k a p p a k = p o − p e 1 − p e , (13) where p o denotes the overall model accuracy, and p e denotes the agreement expected by chance between the model’s predictions and the actual class values (McHugh, 2012).

The stroke risk prediction model was built and trained using the PyTorch. Experiments were conducted on a PC with Intel(R) Xeon(R) Gold 6258R CPU @ 2.70 GHz and NVIDIA QuADro GV100 GPU. We trained the model with the Adam optimizer (Kingma and Ba, 2014) with default parameters and a fixed learning rate of 0.001. And we randomly select 80% of the samples from whole dataset for training, and the remaining 20% for testing. The maximum number of epochs employed for training is 100. The datasets and source codes are publicly available on GitHub: https://github.com/zhangdaoliang/SRPNet.

We utilized a dataset from the CSDC database, consisting of 862,244 samples, with each sample originally having 26 distinct features. The proposed two-level feature selection method was used to screen out significant stroke features, which has a positive effect on improving the performance of the prediction model. In the first step of feature selection, we employed Lasso, elastic net, chi-square test and Pearson correlation methods for the initial screening of stroke-related factors. Here, we consider using α = 0.5 for the elastic net. Figure 3 shows that the impact of different parameters contained in these methods on the feature selection results. We can observe in Figures 3A, B the paths of regression coefficient changes based on Lasso and elastic net, with each curve corresponding to one feature variable. Figures 3C, D demonstrate the correlation of each feature with stroke. It is worth noting that we tend to select features with higher scores and p ≤ 0.05 in the chi-square test (Pandis, 2016). According to Figure 3, Lasso, elastic net, chi-square test and Pearson correlation methods select 13, 15, 8 and 12 features, respectively.

The specific feature selection results of each method are shown in Table 4. Subsequently, we took the union of features selected by the four methods as the robust candidate feature set, which includes 16 features, i.e., AG, Gender, Smoking, MS, Occupation, ES, HS, HYP, AF, LDL-C, Diabetes, LE, Overweight, FHS/HYP/CHD, PLT, and PLH. The receiver operating characteristic (ROC) curves (Fan et al., 2006) corresponding to different feature sets are shown in Figure 4. We find that using the candidate feature set achieves better prediction results than features selected by individual methods. It illustrates that the first step of feature selection is of great significance for stroke risk diagnosis.

In the second step of feature selection, we eliminate risk factors with strong correlation between features. Based on the results of the first step of feature selection, we iterate through all candidate feature combinations. All feature combinations are evaluated under different machine learning methods as classifiers. The optimal feature combinations for different number of features are determined with respect to the evaluation results. Figure 5 shows the performance of the method with different numbers of feature variables. We see that as the number of features increases, the micro precision of most machine learning methods gradually improves and tends to stabilize. However, the performance of the DT and AdaBoost methods decreases significantly when the number of features is 9 and 15 respectively. When the number of features reaches 12, all seven machine learning methods overall achieve the best performance. Finally, we obtained risk factors that are highly relevant to stroke patients and have no redundant information among features, including Smoking, Occupation, ES, HS, HYP, AF, LDL-C, Diabetes, LE, Overweight, FHS/HYP/CHD, and PLT. It is worth noting that traditional methods consider age and gender to be strongly correlated with stroke risk (Howard et al., 2023; Ospel et al., 2023). However, two-level feature selection has removed them due to their redundancy with occupation and other risk factors. In contrast, the PLT features reflecting the climate of the patient’s location are preserved, and it has been confirmed that low temperatures are associated with an increased risk of stroke (Chen et al., 2013). This indicates that SRPNet could provide new insights for future risk screening.

In this section, we validate the effectiveness of the SRPNet model on the CSDC dataset. Decision tree C5.0 (C5.0) (Ahmadi et al., 2018), random forests (RF) (Breiman, 2001),FCN, one-dimensional convolutional neural network (CNN), long short-term memory network (LSTM) and Transformer are used as comparison methods to predict stroke risk. Table 5 shows the prediction performance of the seven methods on the original CSDC data (all features) and the data after two-level feature selection (selected features). We can find that SRPNet model obtains the best prediction results in terms of the four evaluation metrics. The performance of all predictors after two-level feature selection is significantly better than their performance when using all features. This demonstrates that the two-level feature selection can effectively filter weak and redundant information, thus improving the results of all predictors. On the selected feature data, SRPNet outperformes FCN and Transformer by approximately 1.4%, 1.4%, 12% and 3.2% on metrics micro F1-score, micro precision, macro precision, Cohen’s Kappa coefficient. This reflects that deep fusion network can better explore potential relationships between risk factors. In summary, the proposed SRPNet model is reasonable and effective for predicting stroke risk.

Furthermore, to make the results more convincing, we evaluated six predictors on in-house data from affiliated hospital of Jining Medical University. The experimental results are recorded in Table 6. We can draw the similar conclusion that the proposed SRPNet model is an ideal and effective prediction tool of stroke risk. To explore the features that play a dominant role in precise classification, we removed each feature and obtained the prediction results for stroke risk. We found that after removing the hypertension (HYP) feature resulted in micro F1-score, micro precision, macro precision, and Cohen’s Kappa coefficient of 0.7, 0.7, 0.83, and 0.28 respectively, which had the greatest impact on stroke prediction performance. Secondly, gender and age also significantly influenced stroke classification, while they are identified as redundant features in the CSDC dataset. The reason is that the analysis conducted on the CSDC dataset involves complex stroke risk prediction, focusing on differences between multiple risk levels, whereas the in-house dataset only focuses on whether someone has a stroke, conducting a simple stroke prediction analysis. Understanding these risk factors can assist doctors in making quick and accurate stroke diagnoses.

To further evaluate the superiority of SRPNet, we visualize the confusion matrices obtained by the six methods on the CSDC dataset and the in-house dataset in Figures 6, 7, where the columns and rows are the predicted labels and true labels, respectively. It shows that compared to other methods, The SRPNet method wins in all categories in terms of prediction accuracy. Additionally, we discover that the history of stroke (HS) feature and the hypertension (HYP) feature significantly enhance the ability of almost all algorithms in Figure 6 to detect stroke effectively.